IBM Unveils Granite 4.2, Open-Reasoning Models for Local Deployment

IBM has released Granite 4.2, a set of open-reasoning models that can be deployed locally on company hardware. This move gives enterprises the ability to run AI models in-house without relying on paid cloud APIs.

The models are available for download from Hugging Face or GitHub, and can also be run through Ollama. IBM's goal is to provide enterprise buyers with a self-contained solution that doesn't require an API meter.

Granite 4.2 comes in three parameter sizes: 3B, 8B, and 30B. These models are built for agentic workflows across cloud, on-premises, and edge environments, and can handle tasks such as reasoning, tool use, coding, instruction following, and search-driven work.

The training process involved fine-tuning the models on approximately 7.2 million samples, with around 100 billion tokens. The agentic data made up 31.6% of the mixture, with software engineering being the largest slice at 69%. The larger two models went through a more extensive reinforcement learning and RLHF process.

IBM is betting on the appeal of running AI models in-house due to cost savings and data security concerns. With Granite 4.2, companies can fine-tune the models, host them locally, and keep prompts inside their own network.